RESEARCH · RESEARCH · #1104
LAVOIR: single-pass decision encoder that predicts value-of-information to decide when to ask
LAVOIR (Laya with Value-Of-Information Routing) augments Laya-style single-pass decision encoders to output both a decision distribution and, for each candidate missing information slot, an expected value-of-information (VOI) so the system can decide whether to ask a question. The method creates VOI training targets without human labels, uses a Gini-impurity cap to bound predicted VOI, matches a Bayes ceiling on seen schemas, achieves AUC 0.799 vs. a greedy oracle 0.797, improves accuracy by up to 14.1 points with ≤0.5 questions per conversation, and reduces asking on SGD from 93% to 8.6%; median answer latency is 31 ms on GH200.
KEY POINTS
- LAVOIR (Laya with Value-Of-Information Routing) augments Laya-style single-pass decision encoders to output both a decision distribution and, for each candidate missing information slot, an expected value-of-information (VOI) so the system can decide whether to ask a question.
- The method creates VOI training targets without human labels, uses a Gini-impurity cap to bound predicted VOI, matches a Bayes ceiling on seen schemas, achieves AUC 0.799 vs.
- a greedy oracle 0.797, improves accuracy by up to 14.1 points with ≤0.5 questions per conversation, and reduces asking on SGD from 93% to 8.6%; median answer latency is 31 ms on GH200.
WHY IT MATTERS
This shows a practical way to let fast single-pass decision models estimate the value of asking clarifying questions, improving accuracy with few queries and no human-labeled VOI targets.